On Mon, 2025-12-01 at 10:06 -0600, Eric W. Biederman wrote:
Roberto Sassu roberto.sassu@huaweicloud.com writes:
- Mimi, linux-integrity (would be nice if we are in CC when linux-
security-module is in CC).
Apologies for not answering earlier, it seems I don't receive the emails from the linux-security-module mailing list (thanks Serge for letting me know!).
I see two main effects of this patch. First, the bprm_check_security hook implementations will not see bprm->cred populated. That was a problem before we made this patch:
https://patchew.org/linux/20251008113503.2433343-1-roberto.sassu@huaweicloud...
Thanks, that is definitely needed.
Does calling process_measurement(CREDS_CHECK) on only the final file pass review? Do you know of any cases where that will break things?
We intentionally changed the behavior of CREDS_CHECK to be invoked only for the final file. We are monitoring for bug reports, if we receive complains from people that the patch breaks their expectation we will revisit the issue.
Any LSM implementing bprm_check_security looking for brpm->cred would be affected by recalculating the DAC credentials for the final binary.
As it stands I don't think it should be assumed that any LSM has computed it's final creds until bprm_creds_from_file. Not just the uid and gid.
Uhm, I can be wrong, but most LSMs calculate their state change in bprm_creds_for_exec (git grep bprm_creds_for_exec|grep LSM_HOOK_INIT).
If the patch you posted for review works that helps sort that mess out.
Well, it works because we changed the expectation :)
to work around the problem of not calculating the final DAC credentials early enough (well, we actually had to change our CREDS_CHECK hook behavior).
The second, I could not check. If I remember well, unlike the capability LSM, SELinux/Apparmor/SMACK calculate the final credentials based on the first file being executed (thus the script, not the interpreter). Is this patch keeping the same behavior despite preparing the credentials when the final binary is found?
The patch I posted was.
My brain is still reeling from the realization that our security modules have the implicit assumption that it is safe to calculate their security information from shell scripts.
If I'm interpreting this behavior correctly (please any LSM maintainer could comment on it), the intent is just to transition to a different security context where a different set of rules could apply (since we are executing a script).
Imagine if for every script, the security transition is based on the interpreter, it would be hard to differentiate between scripts and associate to the respective processes different security labels.
In the first half of the 90's I remember there was lots of effort to try and make setuid shell scripts and setuid perl scripts work, and the final conclusion was it was a lost cause.
Definitely I lack a lot of context...
Now I look at security_bprm_creds_for_exec and security_bprm_check which both have the implicit assumption that it is indeed safe to compute the credentials from a shell script.
When passing a file descriptor to execat we have BINPRM_FLAGS_PATH_INACCESSIBLE and use /dev/fd/NNN as the filename which reduces some of the races.
However when just plain executing a shell script we pass the filename of the shell script as a command line argument, and expect the shell to open the filename again. This has been a time of check to time of use race for decades, and one of the reasons we don't have setuid shell scripts.
Yes, it would be really nice to fix it!
Yet the IMA implementation (without the above mentioned patch) assumes the final creds will be calculated before security_bprm_check is called, and security_bprm_creds_for_exec busily calculate the final creds.
For some of the security modules I believe anyone can set any label they want on a file and they remain secure (At which point I don't understand the point of having labels on files). I don't believe that is the case for selinux, or in general.
A simple example for SELinux. Suppose that the parent process has type initrc_t, then the SELinux policy configures the following transitions based on the label of the first file executed (sesearch -T -s initrc_t -c process):
type_transition initrc_t NetworkManager_dispatcher_exec_t:process NetworkManager_dispatcher_t; type_transition initrc_t NetworkManager_exec_t:process NetworkManager_t; type_transition initrc_t NetworkManager_initrc_exec_t:process initrc_t; type_transition initrc_t NetworkManager_priv_helper_exec_t:process NetworkManager_priv_helper_t; type_transition initrc_t abrt_dump_oops_exec_t:process abrt_dump_oops_t; type_transition initrc_t abrt_exec_t:process abrt_t; [...]
(there are 747 rules in my system).
If the transition would be based on the interpreter label, it would be hard to express with rules.
If the transition does not occur for any reason the parent process policy would still apply, but maybe it would not have the necessary permissions for the execution of the script.
So just to remove the TOCTOU race the security_bprm_creds_for_exec and security_bprm_check hooks need to be removed, after moving their code into something like security_bprm_creds_from_file.
Or am I missing something and even with the TOCTOU race are setuid shell scripts somehow safe now?
Take this with a looot of salt, if there is a TOCTOU race, the script will be executed with a security context that does not belong to it. But the transition already happened. Not sure if it is safe.
I also don't know how the TOCTOU race could be solved, but I also would like it to be fixed. I'm available to comment on any proposal!
Roberto